Simple self-scalable grid classifier for signal denoising in digital processing systems

This paper presents an approach to build a data classifier based on a simple and inexpensive evaluation function aimed to reduce the computational costs when processing new incoming instances. The classifier agent employs in its training concepts of Self-Organized Maps and Multiple Instance Learning. The motivation for this proposal was the need of a classifier for the processing of signals from partial discharges (PD) in electrical power systems. The monitoring of these signals allows building a database that can be a reference for the detection of failures in equipment and thus support predictive maintenance routines. The simplicity of calculations, which are based on the Manhattan distance (City Block metric), makes it very suitable for use in embedded systems, field-programmable-logic based systems (e. g. FPGAs) and application specific integrated circuits (ASICs). Further improvements can be achieved with pipelining and parallelism techniques, leading to significant maximization of sample rates. The classifier can operate with multiple classes without requiring complex configurations. Other advantages include automatic fitting to the complexity of the problem, the ability to deal with input vectors with different characteristics (inheritance of multiple instance learning) and, depending on implementation, it may be easily upgraded to allow operation with new input patterns (other than those used in the initial training). To evaluate the proposal, we used a set of PD databases obtained by simulations and laboratory measurements. We performed comparisons with standard form classifiers based on Multilayer Perceptrons and Support Vector Machines. The results show that the technique gives the same orders of accuracy and generalization of those classifiers, but with the advantages of low computational costs, self-scalability, dimensional independence and high potential for parallelization.